An Overview of Machine Learning Techniques for Onboard Anomaly Detection in Satellite Telemetry*

Satellite telemetry is often the only insight into how a satellite system is performing. The information contained in the telemetry channels can vary, but all of them can give information about the operation of the spacecraft. When down-linking telemetry, a limiting factor on all space missions is the link budget, i.e., how much data can be downlinked. In nominal spacecraft operation, not all data can be downlinked and is therefore either lost or monitored onboard. Normally, real-time monitoring of several hundred telemetry channels is required for the data that does make it to ground. Monitoring this volume of telemetry data can benefit from some machine assistance for the satellite operators. Typically, this is in the form of basic thresholding or more advanced statistical methods, such as ARIMA, applied to the telemetry signals. However, advances in modern computing hardware and machine learning techniques make onboard solutions feasible. This paper explores using Machine Learning techniques such as Autoencoders, Long Short Term Memory networks, Variational Autoencoders, and Generative Adversarial Networks for onboard anomaly detection from satellite telemetry data. The performance of these methods is compared to basic statistical methods, and the applicability of these models to small-scale Edge AI or FPGA-based hardware systems, allowing deployment onboard spacecraft, is assessed.